The same one-paragraph request was given three times to the same local AI model, through three different agent setups. Every line of code below was written by a model running on a single desktop graphics card. No cloud model wrote any of it.
Build a playable, no copyright harmed Homage to Legend of Zelda 1, playable using HTML5 and JS in the browser. It should be a vertical slice including an overworld and a dungeon, at least two weapons, and three enemies.
A lightweight open-source coding agent with a small system prompt. The first two attempts spent their whole output budget planning the game inside the model's thinking and never wrote a line. A hard cap on thinking fixed that, and the third attempt built the game in one pass. Its art is all canvas rectangles.
WASD/arrows move · Space sword · F bow · B bomb · R restart
Play build 1Anthropic's Claude Code harness pointed at the local model, before we knew what it needed. It finished, but slowly. Claude Code assumed a ~1M-token context window, so it never compacted the conversation and ran into the server's real limit. In auto mode, a permission check before every action kept pushing the main conversation out of the model's cache, so most of the time went to re-reading its own history. Images were off at first, so it inspected its tile sheet as ASCII art.
WASD/arrows move · J/Space sword · K bow · L bomb · E interact · Enter begin
Play build 2The same harness after fixing the setup: compaction tuned to the real context size, auto mode off, and the model's image input turned on. It finished in less than a third of the time and also wrote its own tests: a 163-check harness (map shape, exits, mechanics, solvability) and a bot that plays the full route.
WASD/arrows move · Space attack · 123 weapon · E talk · M sound · R restart
Play build 3